Method for the three-dimensional detection of objects

Abstract

The present invention relates to a method for the three-dimensional detection of objects in which two images of a scene are detected and mutually corresponding objects in the images are identified by means of a correlation process. A binarization scheme which compares randomly selected point pairs with one another is used in the correlation process. The point pairs are fixed by means of a plurality of iterations.

The invention claimed is: 
     
       1. A method for three-dimensional detection of objects, the method comprising the steps of:
 detecting a scene by means of a first image detection device, wherein a first image is produced; 
 detecting the scene by means of a second image detection device, wherein a second image is produced; 
 fixing a correlation window as a portion of the first image; 
 fixing a plurality of search windows as portions of the second image, wherein different search windows reproduce different regions of the second image; 
 determining the one of the plurality of search windows that corresponds to the correlation window by means of a correlation process, 
 wherein 
 in the correlation process, a binarization scheme is applied to the correlation window and to the plurality of search windows in which the respective points of at least one randomly selected point pair within the correlation window and within the plurality of search windows are compared with one another, with a respective results vector being formed for the correlation window and for the plurality of search windows; 
 a correlation of the results vector of the correlation window and of the results vectors of the plurality of search windows takes place to determine the one of the plurality of search windows corresponding to the correlation window; and 
 the at least one point pair is selected by means of a plurality of iterations, wherein the at least one point pair is reselected randomly for each iteration and wherein at least one point pair is used for the correlation process which achieves the best association between the correlation window and the one of the plurality of search windows corresponding to the correlation window. 
 
     
     
       2. The method in accordance with  claim 1 ,
 wherein the at least one point pair is selected before the detection of the scene, with the point pair being the same for the correlation window and for each of the plurality of search windows. 
 
     
     
       3. The method in accordance with  claim 1 ,
 wherein a plurality of point pairs are used which are uniformly distributed within the correlation window. 
 
     
     
       4. The method in accordance with  claim 1 ,
 wherein at least 128 point pairs are selected. 
 
     
     
       5. The method in accordance with  claim 1 ,
 wherein the iterations are carried out on the basis of the first image and of the second image of at least one known scene. 
 
     
     
       6. The method in accordance with  claim 1 ,
 wherein the scene is illuminated with a pattern which is non-self similar at least locally. 
 
     
     
       7. The method in accordance with  claim 6 ,
 wherein the illumination pattern is also varied in the iterations. 
 
     
     
       8. The method in accordance with  claim 1 ,
 wherein the correlation window and the plurality of search windows each comprise the same area. 
 
     
     
       9. The method in accordance with  claim 8 ,
 wherein the same area is smaller than 0.5% of the area of the first and/or second images. 
 
     
     
       10. The method in accordance with  claim 8 ,
 wherein the same area is smaller than 0.1% of the area of the first and/or second images ( 22 ,  24 ). 
 
     
     
       11. The method in accordance with  claim 8 ,
 wherein the same area is smaller than 0.01% of the area of the first and/or second images. 
 
     
     
       12. The method in accordance with  claim 1 ,
 wherein the correlation window and the plurality of search windows are circular. 
 
     
     
       13. The method in accordance with  claim 1 ,
 wherein the correlation of the results vector of the correlation window and of the results vectors of the plurality of search windows takes place with reference to the Hamming distance of the results vectors. 
 
     
     
       14. The method in accordance with  claim 13 ,
 wherein an aggregate is formed over a plurality of Hamming distances and takes place using the aggregate of the correlation. 
 
     
     
       15. The method in accordance with  claim 14 ,
 wherein the aggregate is formed over 3×3 Hamming distances. 
 
     
     
       16. The method in accordance with  claim 1 ,
 wherein the first and second images are rectified and/or are filtered by a noise filter. 
 
     
     
       17. The method in accordance with  claim 1 ,
 wherein the plurality of search windows are only arranged along an epipolar line of the second image. 
 
     
     
       18. A system for three-dimensional detection of objects, the system comprising
 a first image detection device configured to detect a first scene and produce a first image; 
 a second image detection device configured to detect a second scene and produce a second image; and, 
 a computing unit configured to:
 fix a correlation window as a portion of the first image; 
 fix a plurality of search windows as portions of the second image, wherein different search windows reproduce different regions of the second image; 
 determine the one of the plurality of search windows that corresponds to the correlation window by means of a correlation process, 
 wherein 
 in the correlation process, a binarization scheme is applied to the correlation window and to the plurality of search windows in which the respective points of at least one randomly selected point pair within the correlation window and within the plurality of search windows are compared with one another, with a respective results vector being formed for the correlation window and for the plurality of search windows; 
 a correlation of the results vector of the correlation window and of the results vectors of the plurality of search windows takes place to determine the one of the plurality of search windows corresponding to the correlation window; and 
 the at least one point pair is selected by means of a plurality of iterations, wherein the at least one point pair is reselected randomly for each iteration and wherein at least one point pair is used for the correlation process which achieves the best association between the correlation window and the one of the plurality of search windows corresponding to the correlation window.